Convection in the deep Earth dynamically changes the elevation of plates. Amplitudes of those vertical motions predicted from numerical models are significantly higher than observations. We find that at small wavelengths (< 1000 km) this misfit can be due to the oversimplification in viscosity of rocks. By a suite of numerical experiments, we show that considering the non–Newtonian rheology of the mantle results in predictions in amplitude of dynamic topography consistent with observations.